CN112577744B - Rolling bearing fault identification method based on combination of SPA-map and ResNet - Google Patents
Rolling bearing fault identification method based on combination of SPA-map and ResNet Download PDFInfo
- Publication number
- CN112577744B CN112577744B CN202011341214.7A CN202011341214A CN112577744B CN 112577744 B CN112577744 B CN 112577744B CN 202011341214 A CN202011341214 A CN 202011341214A CN 112577744 B CN112577744 B CN 112577744B
- Authority
- CN
- China
- Prior art keywords
- matrix
- trend
- signal
- original
- item
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Abstract
The invention discloses a rolling bearing fault identification method based on the combination of an SPA-map and ResNet, which comprises the steps of firstly utilizing an SPA method to decompose an original signal into a trend item and a trend removing item with larger differences, then combining the obtained trend item and the trend removing item with the original signal item to convert the trend item and the trend removing item into a color map, and finally utilizing a ResNet network model to realize fault type identification, greatly reducing component items under the condition of extracting as much bearing fault information as possible, and simultaneously extracting deep layer bearing fault information through deep ResNet so as to improve the rolling bearing fault identification efficiency and accuracy.
Description
Technical Field
The invention belongs to the technical field of rotary machine fault identification, relates to rolling bearing fault identification, and particularly relates to a rolling bearing fault identification method based on the combination of an SPA-map and a ResNet.
Background
Rolling bearings are one of the most common components in mechanical systems, which play an extremely important role in maintaining the stability of the mechanical system, but are also one of the vulnerable parts. According to the statistics of the faults of the mechanical system, the proportion of the faults of the rolling bearing exceeds 40 percent. When the bearing is damaged, the mechanical system works abnormally, serious accidents are caused seriously, and further economic loss and casualties are caused, so that the accurate fault diagnosis of the bearing is very important to ensure the normal work of the mechanical system. Because the rolling bearing works under complex working conditions, once a fault occurs, the vibration signal of the rolling bearing generally shows the characteristics of non-stability and non-linearity. To realize accurate bearing fault diagnosis, a bearing signal is analyzed, and features which can represent the bearing fault most are extracted from nonlinear and non-stationary vibration signals.
The bearing signal is typically a time series of amplitude variations over time. The traditional fault identification method is to decompose an original bearing time sequence into a plurality of items, then select some of the items as the characteristic information of the rolling bearing, and has the defects that proper components are not easy to select and the number of the components is difficult to determine, if the selection is too much, data repetition and calculation amount increase are caused, and if the selection is too little, the bearing fault characteristics are not completely extracted, so that the final fault diagnosis precision is not high.
In recent years, deep learning is rapidly developed and widely applied to the field of fault mode identification based on data driving, and the method has the greatest characteristic of automatically extracting the features in original data, so that the problems of insufficient feature extraction or inapplicable extraction method are solved. The Convolutional Neural Network (CNN) is one of the widely-used deep learning classical models, and has the advantages of being capable of well extracting image features and being applied to fault diagnosis of a rolling bearing. For example, Eren L and the like directly diagnose original vibration data through one-dimensional CNN, Do V and the like firstly convert a vibration image into a gray scale image, and then recognize the image through two-dimensional CNN to obtain a better result. Chen L et al obtains a high accuracy by reconstructing the time series to obtain a two-dimensional matrix as input to the CNN. Chen Z et al converts the skewness, mean and variance of the original vibration signal into a feature matrix as input to the CNN, diagnosing bearing faults. Although the CNN has a better capability of extracting image features, and as the number of network layers increases, the network can extract deep feature information to obtain a good generalization capability, but when the number of network layers increases, the network becomes more and more difficult to train, because in a network with a large number of layers, when gradient information is transmitted from the last layer to the first layer of the network layer by layer, a phenomenon that the gradient approaches zero or is very large, called gradient dispersion or gradient explosion, occurs, and the deeper the number of network layers, the more serious the phenomenon of gradient dispersion or gradient explosion occurs, and further the deeper the feature is difficult to extract.
Disclosure of Invention
In order to solve the component problem and extract the deep bearing fault characteristic, the invention provides a rolling bearing fault identification method based on the combination of the SPA-map and the ResNet, which greatly reduces the component number under the condition of extracting as much bearing fault information as possible, simplifies the calculation process, avoids component selection and number determination, and can extract the deep bearing fault information through the deep ResNet.
The invention idea is as follows: combining Smooth Priors Analysis (SPA) with a depth Residual network (RseNet), the original time series is first decomposed into two components: a trend item and a trend-removing item; secondly, converting the components into a matrix form, combining the matrix form with an original item conversion matrix to form a time domain map, then automatically extracting deep bearing fault characteristics from the map through ResNet, and finally carrying out fault classification. Researches show that the size of the color atlas also has certain influence on the fault identification accuracy. If the atlas is selected too large, each atlas is large, and the training time and the calculation complexity of the model are increased; if the selection of the maps is too small, the information content of each map is reduced, and therefore the fault identification accuracy is affected. The invention sets the size of the color image asm×mI.e. the side length of the color atlas ismThen need to include in the signal samplem 2Number of signal samples, whereinmThe value range of (2) to (128).
The invention provides a rolling bearing fault identification method based on the combination of an SPA-map and a ResNet, which comprises the following steps:
s1, sampling the rolling bearing fault vibration signal to obtain a plurality of signal samples;
s2, decomposing each acquired rolling bearing signal sample into a trend term and a trend removing term by adopting an SPA method;
s3, for each signal sample, constructing a color image three-channel matrix based on the original items of the signal samples and the corresponding trend items and detrending items to obtain a corresponding color map;
and S4, inputting the obtained color map into the trained ResNet network model, and determining the fault type of the rolling bearing.
In the rolling bearing fault identification method based on the combination of the SPA-map and the ResNet, in step S1, the collected rolling bearing fault vibration signal may be sampled in a continuous sampling manner (i.e., two adjacent signal samples are connected end to end without an overlapping portion) or an overlapping sampling manner (i.e., two adjacent signal samples are partially overlapped with each other). Since the color image is set to a size ofm×mTherefore, each signal sample should containm 2The number of signal sampling points.
In the rolling bearing fault identification method based on the combination of the SPA-map and the ResNet, in step S2, each signal sample is decomposed according to the following method:
for rolling bearing fault vibration signal sampleLUsing SPA method to decompose into trend termsTAnd de-trending itemsDTI.e. by
（1）。
Wherein the trend itemTThe following linear observation model was used:
in the formula (I), the compound is shown in the specification,Hrepresenting an observation matrix;θrepresenting a regression parameter;vindicating an observation error.
At this time, firstly, the method is to findθOf (2) an optimal solution
Then pass through
The trend term of the original data is solved,
the calculation is usually performed by using a least square method, and in the SPA, a regularized least square method is used, that is:
in the formula, λ represents a regularization parameter;D d to representdA matrix of discrete-form expressions of an order differential operator;
setting original dataLIs provided withMLocal extrema points (i.e. number of signal sample points),M=m 2and then the trend term matrix corresponding to the local extremum point is:
the discrete forms of their first and second order trends are:
and (3) sequentially deriving a matrix expressed by the discrete form of the trend of any order as follows:
in the optimization process, the formula (3) needs to be satisfied, namely, a differential term needs to be made
Approaching to 0, solving equation (3) can obtain:
the trend term is then calculated according to equation (8) below:
the detrending term is calculated according to equation (9) below:
observation matrixHThe selection of (1) can be obtained according to the characteristics of the original signal data, and for simple calculation, the inventionHSelecting identity matrices, i.e.
。
The second order differential matrix covers all the first order extreme points, and can better estimate the trend term in the data, so the formula (7) showsd=2, i.e. willD 2Set as a regularization matrix of the form:
then:
thus, by selecting the regularization parameter λ, the original signal can be separatedLTrend item ofTAnd de-trending itemsD。
Based on the principle of the SPA, the components obtained by the SPA decomposition only have a trend term and a trend removing term, and only have a single parameter lambda, so that the extraction process of the bearing fault characteristic information is simplified to a great extent. If the value of lambda is too small, the extraction of the trend item is conservative, and the difference between the trend item and the trend removing item is small, so that the matrix values of a red channel and a green channel are close to each other, and the separability of the state is reduced; when the value of the lambda is too large, the extraction of the trend item is over-excited, the obtained trend item is too stable, and the state separability is also reduced. In the invention, the value range of lambda is 3-8, so that SPA analysis is carried out on the original bearing vibration signal.
In the rolling bearing fault identification method based on the combination of the SPA-map and the ResNet, the step S3 comprises the following sub-steps:
s31, original signal processing, namely, firstly converting original signals of signal samples into original item basic matrixes, and then carrying out normalization processing on the original item basic matrixes to obtain original item normalization matrixes;
s32 trend item signal processing, firstly converting the trend item signal into a trend item basic matrix, and then carrying out normalization processing on the trend item basic matrix to obtain a trend item normalization matrix;
s33 trend item removing signal processing, firstly converting trend item removing signals into a trend item removing basic matrix, and then carrying out normalization processing on the trend item removing basic matrix to obtain a trend item removing normalization matrix;
s34, constructing a three-channel matrix in the color map, and converting the original item normalization matrix, the trend item normalization matrix and the detrended item normalization matrix into three-channel values of the color map to obtain the corresponding color map.
In step S31, the signal sample Original signal is converted into an Original term basis matrix (Original basis matrix,OBM）：
OBM(j, k) base matrix representing original terms of signal samplesjGo to the firstkA matrix element of a column;j=1,2，…，m；k=1,2，…，m；
then, according to equation (13), the Original term normalization matrix (Original normalized matrix,ONM）：
ONM(j,k) Normalized matrix representing original terms of signal samplesjGo to the firstkA matrix element of a column;Max j k,（OBM(j,k) Represents the maximum value in the base matrix of the original entries of the current single signal sample;Min j k,（OBM(j,k) ) represents the minimum value in the base matrix of the original entries of the current single signal sample.
In step S32, the Trend term signal obtained in step S2 is converted into a Trend term basis matrix (Trend basis matrix,TBM）：
TBM(j, k) base matrix number representing trend terms of signal samplesjGo to the firstkA matrix element of a column;j=1,2，…，m；k=1,2，…，m；
then, according to equation (15), the original term normalization matrix (Trend normalized matrix,TNM）：
TNM i (j,k) Normalized matrix expressing signal sample trend termjGo to the firstkA matrix element of a column;Max j k,（TBM(j,k) Represents the maximum value in the current single signal sample trend term basis matrix;Min j k,（TBM(j,k) ) represents the minimum value in the current single signal sample trend term basis matrix.
In step S33, the detrending term signal obtained in step S2 is converted into a detrending term basis matrix (detrended basis matrix,DBM）：
DBM(j, k) base matrix number representing trend terms of de-signal samplesjGo to the firstkA matrix element of a column;j=1,2，…，m；k=1,2，…，m；
then, according to the formula (17), the original term normalization matrix (Dtrend normalized matrix,DNM）：
DNM(j,k) Normalized matrix expressing trend term of de-signal samplejGo to the firstkA matrix element of a column;Max j k,（DBM(j,k) Represents the maximum value in the current single signal sample detrended term basis matrix,Min j k,（DBM(j,k) Is) represents the minimum value in the current single signal sample detrended term basis matrix.
In step S34, the RGB color map is to be constructed, and thus, the original term normalization matrix, the trend term normalization matrix, and the detrended term normalization matrix are multiplied by 255, respectively, to obtain a matrix as three channels in the color map. In the present invention, theONMMultiplied by 255 as the matrix for the first channel (red channel) in the color map, willTNMMultiplied by 255 as a color mapThe second channel (green channel) of the matrix, willDNMMultiplied by 255 as a matrix for the third channel (blue channel) in the color map.
Therefore, through the color map acquisition mode, each constructed channel matrix only containsmAnd one parameter greatly simplifies the map acquisition process.
In step S4, the ResNet network model is the ResNet50 network model.
Compared with the prior art, the rolling bearing fault identification method based on the combination of the SPA-map and the ResNet has the following outstanding advantages and beneficial technical effects:
1. according to the method, firstly, an SPA method is utilized to decompose an original signal into a trend item and a trend removing item with larger differences, the obtained trend item and the trend removing item are combined with the original signal item to be converted into a color map, finally, a ResNet network model is utilized to realize fault type identification, component items are greatly reduced under the condition that as much bearing fault information as possible is extracted, and meanwhile, deep bearing fault information is extracted through deep ResNet, so that the fault identification efficiency and accuracy of the rolling bearing are improved.
2. According to the method, the SPA is used for decomposing the original signal into a trend term and a trend removing term, the trend term and the trend removing term better summarize the characteristics of the original signal and only comprise one regularization parameter, so that the problem of difficulty in component selection is solved; meanwhile, the image conversion process only comprises one image size, so that the preprocessing process of the signals at the early stage is greatly simplified.
3. The method constructs the color map based on the original item signal, the trend item signal obtained by SPA decomposition and the detrending item signal, can contain more characteristic information, is beneficial to subsequent model training, and further improves the accuracy rate of identifying the rolling bearing fault.
4. The identification method can accurately identify 16 classifications of the bearing fault modes, can identify the fault degree and the fault part in different states, and has great guiding significance in practical application.
Drawings
Fig. 1 is a schematic diagram of data partitioning.
FIG. 2 is a schematic diagram of a network model training process in the rolling bearing fault identification method based on the combination of the SPA-map and the ResNet.
FIG. 3 is a schematic diagram of a signal transformation map according to the present invention.
FIG. 4 is a schematic diagram of a process of converting a signal segment into an RGB map according to the present invention.
FIG. 5 is a ResNet network model identification accuracy and loss value variation curve with iteration steps; wherein (a) is a curve of the ResNet network model identification accuracy along with the change of iteration steps aiming at training set, verification set and test set data respectively; (b) the loss values of the ResNet network model are plotted against the number of iteration steps for the training set and the validation set data, respectively.
Fig. 6 is a confusion matrix obtained by classifying the atlas in the embodiment of the present invention.
Fig. 7 is a visualization diagram of image features extracted by a ResNet network model in the embodiment of the present invention.
Detailed Description
The embodiments of the present invention will be given below with reference to the accompanying drawings, and the technical solutions of the present invention will be further clearly and completely described by the embodiments. It is to be understood that the described embodiments are merely exemplary of the invention, and not restrictive of the full scope of the invention. All other embodiments, which can be derived by a person skilled in the art from the disclosure of the invention without any inventive step, are within the scope of the invention.
Example 1
The experimental data used in this example is 6205-2RS JME SKF deep groove ball bearing data (SMITH W, RANDALL r. Rolling element bearing using cement bearing failure data: a benchmark stuck [ J ] Mechanical Systems and Signal Processing,2015,64-65(3): 100-122) in the vibration Signal of Rolling bearing failure opened by the university of kasesque storage (CWRU), and the failure types thereof are divided into inner ring failure, Rolling element failure and outer ring failure, each failure introduces a single point failure by the electric discharge machining method, the failure diameter is 0.007inches, each failure is loaded with 0, 1HP, 2HP and 3HP, and the data for comparison is constructed under the same conditions with a normal bearing, which is specifically shown in table 1.
TABLE 1 16 working conditions under normal and 0.007 fault conditions
When dividing data, firstly, according to a general training set, verification set and test set dividing principle, dividing the data into a training set, a verification set and a test set, wherein the proportion of the training set to the test set of the verification set is set to be 0.6, 0.2 and 0.2.
The ResNet network model used in this embodiment is the ResNet50 network model, and the specific structure can be referred to (He K, Zhang X, Ren S, et al. Deep reactive Learning for image registration [ C ].2016 IEEE Conference on Computer Vision and Pattern registration, 2016: 770-. Since the present embodiment is classified and identified for 16 faults, the output layer of the ResNet50 network model is 16. In this embodiment, the data in the training set and the validation set are first used to train the ResNet network model, as shown in fig. 2, which specifically includes the following steps:
a1 sampling the vibration signal of rolling bearing fault to obtain several signal samples, each of which containsm 2The number of signal sampling points.
The rolling bearing fault vibration signal samples obtained by overlapping sampling in the training set and the verification set are used.
In order to prevent overfitting caused by too little training data, the embodiment further samples the training set and the verification set by using an overlapped sampling manner, that is, by using a fixed-step-length, fixed-length, sliding sampling manner, and the method is shown in fig. 1. Since the color map used in the present embodiment is an RGB map, the map size is selected to be 64 × 64, i.e., the map sizem=64, therefore, resampling the training and validation setsThe number of signal sampling points included in each signal sample obtained by sampling is 4096.
The rolling bearing fault vibration signal samples obtained by overlapping sampling in the training set and the verification set are used.
As shown in fig. 3, next, a color map is generated from the rolling bearing failure diagnosis signal by steps a2 and A3.
A2 decomposes each signal sample in the training set and the verification set into a trend term and a de-trend term by using the SPA method.
For example, for the training setiA signal sampleL i Decomposition into trend terms by SPA methodT i And de-trending itemsDT i I.e. by
，i=1,2，…，N，NRepresenting the total number of signal samples in the training set.
According to the formulae (1) to (11) given above, the first result isiTrend term of individual signal samplesT i And de-trending itemsDT i ：
The value of λ in this embodiment is 5.
Observation matrix in this embodimentHSelecting identity matrices, i.e.
。
Second order differential matrixD 2 Comprises the following steps:
then:
a3, for each signal sample in the training set and the verification set, constructing a color image three-channel matrix based on the original items of the signal samples and the corresponding trend items and detrending items to obtain a corresponding color map.
In this step, the training set isiA signal sampleL i For example, the image conversion process is explained in detail. As shown in fig. 4, in combination with the color map construction process given above, the steps specifically include the following sub-steps:
a31 original signal processing, firstly converting the original signal of the signal sample into an original item basis matrix, and then carrying out normalization processing on the original item basis matrix to obtain an original item normalization matrix.
The signal samples raw signal is converted into a raw term basis matrix (Original basis matrix,OBM）：
OBM i (j, k) representsiFirst of the basis matrix of the original terms of the signal samplesjGo to the firstkThe matrix elements of the columns are arranged in a matrix,i=1,2,…,N；j=1,2，…，m；k=1,2，…，m；
then, according to the formula (13-1), the Original term normalized matrix (Original normalized matrix,ONM）：
ONM i (j,k) Is shown asiNormalization matrix number of original items of signal samplesjGo to the firstkA matrix element of a column;Max j k,（OBM i (j,k) Represents the maximum value in the base matrix of the original entries of the current single signal sample;Min j k,（OBM i (j,k) ) represents the minimum value in the base matrix of the original entries of the current single signal sample.
And A32 trend item signal processing, firstly converting the trend item signal into a trend item basic matrix, and then carrying out normalization processing on the trend item basic matrix to obtain a trend item normalization matrix.
The Trend term signal obtained in step a2 is converted into a Trend term basis matrix (Trend basis matrix,TBM）：
TBM i is shown asiFundamental matrix of individual signal sample trend termsjGo to the firstkA matrix element of a column;i=1,2,…,N；j=1,2，…，m；k=1,2，…，m；
then, according to the formula (15-1), the original term normalization matrix (Trend normalized matrix,TNM）：
TNM i (j,k) Is shown asiNormalization matrix of signal sample trend termjGo to the firstkA matrix element of a column;Max j k,（TBM i (j,k) Represents the maximum value in the current single signal sample trend term basis matrix;Min j k,（TBM i (j,k) ) represents the minimum value in the current single signal sample trend term basis matrix.
A33 detrending item signal processing, firstly converting detrending item signals into detrending item basis matrixes, and then carrying out normalization processing on the detrending item basis matrixes to obtain detrending item normalization matrixes.
The detrending term signal obtained in step a2 is converted into a detrending term basis matrix (detrended basis matrix,DBM）：
DBM i (j, k) representsiBase matrix of detrended terms for individual signal samplesjGo to the firstkThe matrix elements of the columns are arranged in a matrix,i=1,2,…,N；j=1,2，…，m；k=1,2，…，m；
then, according to the formula (17-1), the original term normalization matrix (Dtrend normalized matrix,DNM）：
DNM(j,k) Normalized matrix expressing trend term of de-signal samplejGo to the firstkA matrix element of a column;Max j k,（DBM(j,k) Represents the maximum value in the current single signal sample detrended term basis matrix,Min j k,（DBM(j,k) Is) represents the minimum value in the current single signal sample detrended term basis matrix.
A34, constructing a three-channel matrix in the color map, and converting the original item normalization matrix, the trend item normalization matrix and the detrending item normalization matrix into three-channel values of the color map to obtain the corresponding color map.
In this step, the RGB color map is to be constructed, and therefore, the original item normalization matrix, the trend item normalization matrix, and the detrended item normalization matrix are multiplied by 255 respectively to obtain a matrix as three channels in the color map, which is specifically as follows:
i.e. byONMMultiplied by 255 as the matrix for the first channel (red channel) in the color map, willTNMMultiplied by 255 as a matrix for the second channel (green channel) in the color map, willDNMMultiplied by 255 as a matrix for the third channel (blue channel) in the color map.
And then, labeling each color image to be used as input data of a ResNet network model, wherein each label calibration value corresponds to a rolling bearing fault category.
Fig. 3 shows the processing results obtained after the signal sample in the training set is processed according to the above steps a2 and A3, and it can be seen from fig. 3 that the decomposed trend term and detrending term are both greatly different from the original signal, and there is an obvious difference between the two terms, which indicates that the two terms contain different fault information, thereby indicating that the SPA can effectively extract fault feature information.
A4 inputs the color atlas in the training set obtained in the step A3 into a ResNet network model to train the model.
The set ResNet network model is first initialized. In this embodiment, an initial value of a model parameter w of the ResNet network model is set, a crossEntropyLoss function is used as a loss function, the number of samples selected by one training, batch _ size, is 64, the threshold of iteration steps is 20, the learning rate is 0.0001, and the random inactivation probability is 0.5.
Then, the data is input into the ResNet50 network model in batch mode (the number of samples per batch is 64), and the error between the model output value and the label calibration value, namely the loss value, is calculated by using the loss function.
A5 judging whether the ResNet network model is converged, if the model is converged, finishing the training of the ResNet network model; if the model does not converge, step A6 is entered.
In this embodiment, whether the model converges is determined according to the iteration step number, and if the iteration step number does not reach the set iteration step number threshold, the iteration step number increases by 1 and then the process proceeds to step a6 until the iteration step number reaches the threshold.
A6 optimizes the ResNey network model and returns to step A4.
The present embodiment updates the ResNet network model parameters layer by layer using optimizer back-propagation based on the loss values.
In this embodiment, the optimizer employs an Adam optimizer.
The Adam optimizer has optimal performance, can adaptively adjust the learning rate, stores the square exponential average value of the historical gradient and maintains the exponential decay average value of the historical gradient. The updating method is as follows:
in the formula:urepresents the gradient first moment (mean), upsilon represents the gradient second moment (variance), beta1、β2Represents the attenuation coefficient and has a value range of [0,1 ]]，lRepresenting the current number of iteration steps.
The parameter update formula obtained by applying the Adam optimizer is as follows:
in the formula:αthe learning rate is represented, and the value is 0.0001;εrepresents a very small constant, preventing the denominator from being zero;w l is shown aslModel parameter values used in the iterative process.w l+1Represents the second result of optimization with Adam optimizerlModel parameter values in +1 iterations.
In the training process, as the number of training iteration steps increases, the accuracy and loss value obtained by the training set input model change as shown in fig. 5.
Meanwhile, after each iteration, the color map data in the verification set is input into the ResNet network model to obtain the corresponding accuracy and loss value, so that the state of the network model is observed. With the increase of the training iteration steps and the continuous learning of the model, the accuracy and the loss value corresponding to the verification set are changed as shown in fig. 5.
As can be seen in fig. 5, both the training set and the validation set converge rapidly between generations 5-10, and the network remains stable in subsequent iterations.
And further carrying out fault classification on the test set data by using the trained ResNet network model so as to verify the classification effect of the rolling bearing fault identification method.
And (3) carrying out rolling bearing fault identification on the test set data by using the trained ResNet network model according to the following steps:
s1, sampling the rolling bearing fault vibration signal to obtain a plurality of signal samples, wherein each signal sample contains m2The number of signal sampling points.
The rolling bearing fault vibration signal samples obtained by overlapping sampling in the test set are used.
In order to prevent overfitting caused by too little training data, the present embodiment further samples the test set by using an overlapped sampling manner, that is, by using a fixed-step, fixed-length, sliding sampling manner, which is shown in fig. 1. Since the color map used in the present embodiment is an RGB map, the map size is selected to be 64 × 64, i.e., the map sizem=64, so when resampling the test set, the number of signal sample points per signal sample sampled is 4096.
S2 decomposes each signal sample in the test set into a trend term and a detrended term using the SPA method.
E.g. for test set number onei′A signal sampleL i′ Decomposition into trend terms by SPA methodT i′ And de-trending itemsDT i′ I.e. by
，i′=1,2，…，N′，N′Representing the total number of signal samples in the test set.
According to the formulae (1) to (11) given above, the first result isi' Trend term of signal sampleT i′ And de-trending itemsDT i′ ：
The value of λ in this embodiment is 5.
Observation matrix in this embodimentHSelecting identity matrices, i.e.
。
Second order differential matrixD 2 Comprises the following steps:
then:
s3, for each signal sample in the test set, constructing a color image three-channel matrix based on the original items of the signal samples and the corresponding trend items and detrending items to obtain a corresponding color map, wherein the method comprises the following steps:
and S31, original signal processing, namely, firstly converting the original signal of the signal sample into an original item basic matrix, and then carrying out normalization processing on the original item basic matrix to obtain an original item normalization matrix.
The signal samples raw signal is converted into a raw term basis matrix (Original basis matrix,OBM）：
OBM i′ (j, k) representsi′First of the basis matrix of the original terms of the signal samplesjGo to the firstkThe matrix elements of the columns are arranged in a matrix,i′=1,2,…,N′；j=1,2，…，m；k=1,2，…，m；
then, according to the formula (13-2), the Original term normalization matrix (Original normal normalized matrix) is obtained from the maximum value and the minimum value of the single samplezed matrix，ONM）：
ONM i′ (j,k) Is shown asi′Normalization matrix number of original items of signal samplesjGo to the firstkA matrix element of a column;Max j k,（OBM i′ (j,k) Represents the maximum value in the base matrix of the original entries of the current single signal sample;Min j k,（OBM i′ (j,k) ) represents the minimum value in the base matrix of the original entries of the current single signal sample.
And S32 trend item signal processing, namely converting the trend item signal into a trend item basic matrix, and then carrying out normalization processing on the trend item basic matrix to obtain a trend item normalization matrix.
The Trend term signal obtained in step S2 is converted into a Trend term basis matrix (Trend basis matrix,TBM）：
TBM i′ is shown asi′Fundamental matrix of individual signal sample trend termsjGo to the firstkA matrix element of a column;i′=1,2,…,N′；j=1,2，…，m；k=1,2，…，m；
then, according to the formula (15-2), the original term normalization matrix (Trend normalized matrix,TNM）：
TNM i′ (j,k) Is shown asi′Normalization matrix of signal sample trend termjGo to the firstkA matrix element of a column;Max j k,（TBM i′ (j,k) Represents the maximum value in the current single signal sample trend term basis matrix;Min j k,（TBM i′ (j,k) ) represents the minimum value in the current single signal sample trend term basis matrix.
And S33, processing the detrending item signal, firstly converting the detrending item signal into a detrending item basic matrix, and then carrying out normalization processing on the detrending item basic matrix to obtain a detrending item normalization matrix.
The detrending term signal obtained in step S2 is converted into a detrending term basis matrix (detrended basis matrix,DBM）：
DBM i′ (j, k) representsi′Base matrix of detrended terms for individual signal samplesjGo to the firstkThe matrix elements of the columns are arranged in a matrix,i′=1,2,…,N′；j=1,2，…，m；k=1,2，…，m；
then, according to the formula (17-2), the original term normalization matrix (Dtrend normalized matrix,DNM）：
DNM i′ (j,k) Is shown asi′Normalization matrix of individual signal sample detrending itemjGo to the firstkA matrix element of a column;Max j k,（DBM i′ (j,k) Represents the maximum value in the current single signal sample detrended term basis matrix,Min j k,（DBM i′ (j,k) Is) represents the minimum value in the current single signal sample detrended term basis matrix.
S34, constructing a three-channel matrix in the color map, and converting the original item normalization matrix, the trend item normalization matrix and the detrended item normalization matrix into three-channel values of the color map to obtain the corresponding color map.
In this step, the RGB color map is to be constructed, and therefore, the original item normalization matrix, the trend item normalization matrix, and the detrended item normalization matrix are multiplied by 255 respectively to obtain a matrix as three channels in the color map, which is specifically as follows:
i.e. byONMMultiplied by 255 as the matrix for the first channel (red channel) in the color map, willTNMMultiplied by 255 as a matrix for the second channel (green channel) in the color map, willDNMMultiplied by 255 as a matrix for the third channel (blue channel) in the color map.
The resulting color image is then used directly as input data for the ResNet network model.
And S4, inputting the color map in the test set acquired in the step S3 into the trained ResNet network model, and determining the fault type of the rolling bearing.
And inputting the color atlas in the test set into the trained ResNet network model, outputting a fault class label, and determining a corresponding fault class according to the fault class label. By the rolling bearing fault identification method provided by the embodiment, the accuracy rate corresponding to the test set is 99.7%.
The confusion matrix obtained by identifying and classifying the test set is shown in fig. 6, wherein the ordinate represents the actual sample fault condition, and the abscissa represents the predicted fault condition of the model. The numbers in the boxes represent percentages, the 8 th row and the 8 th column of the confusion matrix are taken as examples, the actual working condition is 8, the numbers in the boxes are 99, the graph representing that the model predicts 99% of the graph as the working condition 8, the 9 th column and the 9 th row of the 8 th row are 1, the graph representing that the model predicts the graph with the actual 1% of the working condition 8 as the working condition 9, the rest numbers in the 8 th row are 0, the model does not predict the rest working conditions, and the rest numbers are analogized. It can be seen from the confusion matrix of fig. 6 that the other conditions are 100% accurate except that the predictions of conditions 1, 2, 8, and 9 are within 2% error. The classification result presented by the confusion matrix can directly and clearly see the classification accuracy of each working condition in percentage form from the graph, so that each working condition can be conveniently analyzed.
And (3) reducing the dimension of the image features extracted from the previous layer of the Resnet network model output layer to obtain a two-dimensional plane visual image, as shown in FIG. 7. Different colors in the diagram represent different fault working conditions, and the larger the distance between the color points of different colors is, the more obvious the difference between the color points of different colors is, namely, the separability is high; the smaller the distance between the same color points is, the stronger the characteristic of the feature extraction is, i.e. the clustering performance is good. It can be seen from the figure that the rolling bearing fault identification method provided by the embodiment can better classify various working conditions, and corresponds to 99.70% of accuracy in an experiment.
In order to prove 2 the advantages of the SPA-map and ResNet combined rolling bearing fault identification method provided by the present invention, the present embodiment further compares the Support Vector Machine (SVM), the Random Forest (RF), the Nearest neighbor classification algorithm (K-Nearest neighbor, KNN), the shallow CNN (four layers), and the Deep Neural Network (DNN).
The SVM model is proposed by Vapnik to solve the problemA class of machine learning methods for linear and high-dimensional pattern recognition is based on statistical learning theory and structural risk minimization principle, and seeks a compromise between model complexity and learning ability according to limited sample information so as to obtain the best generalization ability (cortex, Corinna, and Vladimir Vapnik. "Support vector machine.)"Machine learning 20.3 (1995): 273-297）。
The RF model is an algorithm integrating a plurality of trees by the idea of ensemble learning, the basic unit of the RF model is a decision tree, and the RF model essentially belongs to a large branch of machine learning, namely, the ensemble learning method (Breiman, Leo.Machine learning 45.1 (2001): 5-32）。
The DNN network model is a simpler network type in deep learning, and is formed by stacking a plurality of linear or nonlinear neural networks, wherein the network only comprises a full-connection layer, and the DNN network model adopts a four-layer network for comparative analysis.
The SVM model, the RF model and the DNN network model are respectively trained by adopting the color map data in the training set and the verification set of the embodiment, and the fault recognition effect of the trained SVM model, RF model and DNN network model is tested by utilizing the color map data in the test set. The data in table 1 were used for the experiments and the comparison results are shown in table 2.
It can be seen from table 2 that the accuracy achieved by the conventional machine learning method is not high, and in contrast, the accuracy achieved by the deep learning method can be more than 90%, which indicates that the learning capability of deep learning is stronger than that of machine learning, but the recognition method provided by this embodiment has higher accuracy than the two deep learning methods, namely CNN and DNN, because only shallow fault features are extracted by CNN and DNN, and the superiority and feasibility of the method provided by the present invention are further verified by comparison experiments.
TABLE 2 comparative experiments with different algorithms
Example 2
In order to further observe the classification condition of the fault part in the rolling bearing fault identification method provided by the invention, 2 groups of 16 classification experiments are carried out according to the method given in embodiment 1 by selecting data with damage degrees of 0.014inches and 0.021inches respectively (specific working conditions are shown in table 4). Firstly, data are divided into a training set, a verification set and a test set, then a ResNet network model (the structure of the ResNet network model is the same as that in the embodiment 1) is trained according to the steps A1-A6 by using the data in the training set and the verification set, then the trained ResNet network model is tested according to the steps S1-S4 by using the data in the test set, and the experimental result under 0.007inches are recorded in a table 3. As can be seen from Table 3, the accuracy rates under the three damage degrees are all more than 99%, which indicates that the method can effectively identify the fault part of the bearing under different loads.
TABLE 3 Experimental results for three damage degrees
Example 3
In order to further observe the performance of the fault degree classification of the rolling bearing fault identification method provided by the invention, all normal data and all damage degrees of an inner ring, an outer ring (6 o 'clock direction) and a rolling body under all loads are selected, and each group of 16 data is subjected to 3 groups of experiments, wherein the working condition of the inner ring is shown in table 4, and the working conditions of the outer ring (6 o' clock direction) and the rolling body are similar to those in table 4. The data is first divided into a training set, a verification set and a test set, a ResNet network model (having the same structure as that in example 1) is then trained according to steps A1-A6 by using the data in the training set and the verification set, the trained ResNet network model is then tested according to steps S1-S4 by using the data in the test set, and the experimental results are recorded in Table 5.
TABLE 4 16 normal and inner ring operating conditions
TABLE 5 Experimental results at three injury sites
As can be seen from table 5, the faults of the inner ring, the rolling body and the outer ring are respectively 99.93%, 99.70% and 99.97%, and the accuracy rates of the three are all above 99%, which indicates that the fault identification method for the rolling bearing provided by the invention can effectively realize the classification of the fault degree.
In summary, the rolling bearing fault identification method based on the combination of the SPA-map and the ResNet provided by the embodiment diagnoses the bearing fault by combining the SPA and the ResNet, can greatly reduce component items while keeping rich information of an original signal, thereby improving the identification efficiency while considering the accuracy of the rolling bearing fault identification, has strong applicability, can be used for various different working conditions (including inner ring fault, rolling element fault, outer ring fault, damage degree, load and the like), and realizes accurate classification of the bearing fault mode.
Claims (7)
1. A rolling bearing fault identification method based on combination of an SPA-map and a ResNet is characterized by comprising the following steps:
s1, sampling the rolling bearing fault vibration signal to obtain a plurality of signal samples, wherein each signal sample should containm 2The number of sampling points of each signal is,mthe side length of the color atlas is shown;
s2, decomposing each acquired rolling bearing signal sample into a trend term and a trend removing term by adopting an SPA method;
s3, for each signal sample, constructing a color image three-channel matrix based on the original items of the signal samples and the corresponding trend items and detrending items to obtain a corresponding color map; the method comprises the following steps:
s31, original signal processing, namely, firstly converting original signals of signal samples into original item basic matrixes, and then carrying out normalization processing on the original item basic matrixes to obtain original item normalization matrixes;
the method specifically comprises the following steps: converting the signal sample Original signal into an Original term basis matrix Original according to a formula (12),OBM：
OBM(j, k) base matrix representing original terms of signal samplesjGo to the firstkA matrix element of a column; j=1,2，…，m；k=1,2，…，m；Lrepresenting the original signal;
then according to the formula (13), the Original term normalized matrix is obtained from the maximum value and the minimum value of the single sample,ONM：
ONM(j,k) Normalized matrix representing original terms of signal samplesjGo to the firstkA matrix element of a column;Max j k,（OBM(j,k) Represents the maximum value in the base matrix of the original entries of the current single signal sample;Min j k,（OBM(j,k) Represents the minimum value in the current single signal sample original entry basis matrix;
s32 trend item signal processing, firstly converting the trend item signal into a trend item basic matrix, and then carrying out normalization processing on the trend item basic matrix to obtain a trend item normalization matrix;
the method specifically comprises the following steps: the Trend term signal obtained in step S2 is converted into a Trend term basis matrix Trend base matrix according to the following formula,TBM：
TBM(j, k) represents signal samplesFundamental matrix of the present trend termjGo to the firstkA matrix element of a column;j=1,2，…，m；k=1,2，…，m；
then according to the formula (15), a Trend term normalization matrix Trend normalized matrix is obtained from the maximum value and the minimum value of the single sample,TNM：
TNM i (j,k) Normalized matrix expressing signal sample trend termjGo to the firstkA matrix element of a column;Max j k,（TBM(j,k) Represents the maximum value in the current single signal sample trend term basis matrix;Min j k,（TBM(j,k) Represents the minimum value in the current single signal sample trend term basis matrix;
s33 trend item removing signal processing, firstly converting trend item removing signals into a trend item removing basic matrix, and then carrying out normalization processing on the trend item removing basic matrix to obtain a trend item removing normalization matrix;
the detrending term signal obtained in step S2 is converted into a detrending term basis matrix according to the following formula,DBM：
DBM(j, k) base matrix number representing trend terms of de-signal samplesjGo to the firstkA matrix element of a column; j=1,2，…，m；k=1,2，…，m；
then, according to the formula (17), the normalization matrix Dtrend normalized matrix is obtained from the maximum value and the minimum value of the single sample,DNM：
DNM(j,k) Normalized matrix expressing trend term of de-signal samplejGo to the firstkA matrix element of a column;Max j k,（DBM(j,k) Represents the maximum value in the current single signal sample detrended term basis matrix,Min j k,（DBM(j,k) Represents the minimum value in the base matrix of the detrended term for the current single signal sample;
s34, constructing a three-channel matrix in the color map, and converting the original item normalization matrix, the trend item normalization matrix and the de-trend item normalization matrix into three-channel values of the color map to obtain the corresponding color map;
and S4, inputting the obtained color map into the trained ResNet network model, and determining the fault type of the rolling bearing.
2. The rolling bearing fault identification method based on the combination of the SPA-map and the ResNet according to claim 1, characterized in that the collected rolling bearing fault vibration signals are sampled in a continuous sampling mode or an overlapping sampling mode.
3. The method for identifying the fault of the rolling bearing based on the combination of the SPA-map and the ResNet according to claim 1, wherein in the step S2, each signal sample is decomposed according to the following method:
for rolling bearing fault vibration signal sampleLUsing SPA method to decompose into trend termsTAnd de-trending itemsDI.e. by
；
In the formula (I), the compound is shown in the specification,His a matrix of the units,D d a matrix representing a discrete form expression of an arbitrary order trend, and λ represents a set regularization parameter.
4. The method for identifying the fault of the rolling bearing based on the combination of the SPA-map and the ResNet according to claim 3, characterized in that the fault of the rolling bearing is identifiedD d Taking a second order differential matrix, i.e.d=2，
And will beD 2Set as a regularization matrix of the form:
5. the rolling bearing fault identification method based on the combination of the SPA-map and the ResNet according to claim 1, wherein in the step S34, the color map to be constructed is an RGB color map.
6. The method for identifying the fault of the rolling bearing based on the combination of the SPA-map and the ResNet according to claim 5, wherein the original term normalization matrix, the trend term normalization matrix and the detrended term normalization matrix are respectively multiplied by 255 to obtain a matrix which is used as three channels in the color map.
7. The method for identifying the fault of the rolling bearing based on the combination of the SPA-map and the ResNet according to claim 1, wherein in the step S4, a ResNet50 network model is adopted as the ResNet network model.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2020112014621 | 2020-11-02 | ||
| CN202011201462 | 2020-11-02 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112577744A CN112577744A (en) | 2021-03-30 |
| CN112577744B true CN112577744B (en) | 2021-12-03 |
Family
ID=75123492
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011341214.7A Active CN112577744B (en) | 2020-11-02 | 2020-11-25 | Rolling bearing fault identification method based on combination of SPA-map and ResNet |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112577744B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113591654B (en) * | 2021-07-22 | 2023-09-01 | 中南大学 | Zinc flotation working condition identification method based on long-time-range depth characteristics |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107560851A (en) * | 2017-08-28 | 2018-01-09 | 合肥工业大学 | Rolling bearing Weak fault feature early stage extracting method |
| CN110321963A (en) * | 2019-07-09 | 2019-10-11 | 西安电子科技大学 | Based on the hyperspectral image classification method for merging multiple dimensioned multidimensional sky spectrum signature |
| CN110702411A (en) * | 2019-09-23 | 2020-01-17 | 武汉理工大学 | Residual error network rolling bearing fault diagnosis method based on time-frequency analysis |
| CN111797567A (en) * | 2020-06-09 | 2020-10-20 | 合肥工业大学 | Deep learning network-based bearing fault classification method and system |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11481583B2 (en) * | 2017-12-28 | 2022-10-25 | Intel Corporation | Algorithm management blockchain |
-
2020
- 2020-11-25 CN CN202011341214.7A patent/CN112577744B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107560851A (en) * | 2017-08-28 | 2018-01-09 | 合肥工业大学 | Rolling bearing Weak fault feature early stage extracting method |
| CN110321963A (en) * | 2019-07-09 | 2019-10-11 | 西安电子科技大学 | Based on the hyperspectral image classification method for merging multiple dimensioned multidimensional sky spectrum signature |
| CN110702411A (en) * | 2019-09-23 | 2020-01-17 | 武汉理工大学 | Residual error network rolling bearing fault diagnosis method based on time-frequency analysis |
| CN111797567A (en) * | 2020-06-09 | 2020-10-20 | 合肥工业大学 | Deep learning network-based bearing fault classification method and system |
Non-Patent Citations (2)
| Title |
|---|
| 基于平滑先验分析和模糊熵的滚动轴承故障诊断;戴邵武 等;《航空动力学报》;20191031;第34卷(第10期);第2218-2225页 * |
| 结合时频分析和卷积神经网络的滚动轴承故障诊断优化方法研究;黄驰城;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20190515(第5期);C029-204 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112577744A (en) | 2021-03-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Kumar et al. | Disease detection in coffee plants using convolutional neural network | |
| Panigrahi et al. | A cnn approach for corn leaves disease detection to support digital agricultural system | |
| CN111368885B (en) | Gas circuit fault diagnosis method for aircraft engine | |
| Radovanović et al. | Image-based plant disease detection: a comparison of deep learning and classical machine learning algorithms | |
| CN111000553B (en) | Intelligent classification method for electrocardiogram data based on voting ensemble learning | |
| CN113203566B (en) | Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN | |
| CN113486578B (en) | Method for predicting residual life of equipment in industrial process | |
| Anim-Ayeko et al. | Automatic blight disease detection in potato (Solanum tuberosum L.) and tomato (Solanum lycopersicum, L. 1753) plants using deep learning | |
| CN115221930B (en) | Fault diagnosis method for rolling bearing | |
| CN113033309A (en) | Fault diagnosis method based on signal downsampling and one-dimensional convolution neural network | |
| CN112577744B (en) | Rolling bearing fault identification method based on combination of SPA-map and ResNet | |
| Qiao et al. | Detection and classification of early decay on blueberry based on improved deep residual 3D convolutional neural network in hyperspectral images | |
| CN115879026A (en) | Rotary drilling rig hydraulic system fault diagnosis method based on convolutional neural network-integrated classifier | |
| Borugadda et al. | Transfer Learning VGG16 Model for Classification of Tomato Plant Leaf Diseases: A Novel Approach for Multi-Level Dimensional Reduction. | |
| CN116894215B (en) | Gear box fault diagnosis method based on semi-supervised dynamic graph attention | |
| Paliwal et al. | An Overview of Deep Learning Models for Foliar Disease Detection in Maize Crop | |
| CN116644378A (en) | Ultra-long multidimensional time sequence classification method based on double-layer nested LSTM | |
| CN114004258B (en) | Semi-supervised electrocardiographic abnormality detection method | |
| Tripathy et al. | Image processing–based artificial intelligence system for rapid detection of plant diseases | |
| Jia et al. | COVID-19 diagnosis from CT images with convolutional neural network optimized by marine predator optimization algorithm | |
| CN115661739A (en) | Vineyard pest fine-grained identification method based on attribute characteristic knowledge graph | |
| CN114549962A (en) | Garden plant leaf disease classification method | |
| CN112683532A (en) | Cross-working condition countermeasure diagnostic method for bearing | |
| Almola et al. | Citrus diseases recognition by using CNN | |
| Kishore et al. | Paddy Leaf Disease Detection using Deep Learning Methods |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |